AGENT SKILL

using-pdpowers

Routing skill for writing, revising, walking, fact-checking and checking an NSF Project Description (and its Project Summary). Use for "write the project description", "revise the proposal", "walk the description with me", "fact-check the proposal", "write the prior support / broader impacts / AI statement", "fill the remaining pages", or any request about NSF proposal prose. Routes each request to one pd-* leaf skill or to the grant checkers; the leaves encode how proposal writing differs from manuscript writing.

License 许可
GPL-3.0
最近更新
2026-11-13
Stars
★ 0
Forks
0

01 前言 Front Matter

name
using-pdpowers
description
Routing skill for writing, revising, walking, fact-checking and checking an NSF Project Description (and its Project Summary). Use for "write the project description", "revise the proposal", "walk the description with me", "fact-check the proposal", "write the prior support / broader impacts / AI statement", "fill the remaining pages", or any request about NSF proposal prose. Routes each request to one pd-* leaf skill or to the grant checkers; the leaves encode how proposal writing differs from manuscript writing.

02 资源文件 2

03 SKILL.md 正文

using-pdpowers

A Project Description is read once, fast, by a panel mostly outside the subject, and judged on what it will produce, whether the PI can do it, and whether every claim can be trusted. Announce the route, then follow the leaf; leaves do not re-route.

Contracts and precedence

Before any edit, locate the proposal folder and read, in order:

  1. The PI's current instructions in the conversation.
  2. The grants workspace's decision records and contracts when they exist (CONTEXT.md, docs/adr/, GRANT-STYLE.md, GRANT-FORMAT.md).
  3. The proposal folder's REVIEW.md, COMMENTS.md and reviews/.
  4. The solicitation and the PAPPG chapter it defers to. Check page limits there; do not take them from memory.

The PI's instructions win; then the contracts; then these skills. No leaf states a result the evidence does not support, whatever the instruction.

Dispatch

Situation Route
New proposal folder from a template create-nsf-grant (scaffold), then the leaves below per section
"Walk the description with me", "show me the next section", PI annotations pd-section-walk
"Is this true?", claim/number/citation/pinpoint checks, "check the dates" pd-fact-check — run it before asking the PI anything a file or paper can answer
Decisions the draft depends on (headline, scope, what to cut) grill-with-docs, recording ADRs and glossary terms; then return here
Overview / headline / solicitation framing pd-overview
Background, definitions for the panel, prior literature, classical anchor pd-background
Worked example, table of real records pd-data-table (inside the section that needs it)
Software and AI assistance statement for the proposal itself pd-ai-statement
Results from Prior NSF Support pd-prior-support
One proposed project (foundation, objective, preliminary results, plan, feasibility, risks) pd-project
Method, tools, orchestration, records, due diligence pd-method
Broader Impacts pd-broader-impacts
Project Summary write last, from the finished description: pd-overview rules, one page, three headed parts, third person
Reviewer-perspective pass (one subagent per role or named reviewer) pd-review-as; reviewers from gp-potential-reviewers
Apply a comment log or review tables to the draft pd-apply-feedback
“Is this paragraph good?” / paragraph-by-paragraph pass over the description check-paragraph — one subagent per paragraph via its split_paragraphs.py, reader = NSF panelist outside the subject; verdicts to reviews/paragraphs-<file>-<date>.md, no edits
Policy audit (references/PD-POLICY.md, P1–P17) python3 scripts/check_pd_policy.py <description.tex> (mechanical rules), then the [J] rules by hand
Style / format audit (pages, fonts, headings, URLs, build) check-grant-style / check-grant-format
"Sounds like AI", "too many claudeisms", de-slop (P17) purge-claudisms — scan, rewrite hits in place, rescan to zero; before the style audit
Mathematics that needs proving or a research gap using-mathpowers, then return here
Any other proposal document (references, data plan, facilities, budget and justification, COA, biosketch, current and pending, mentoring), Research.gov, UVMClick, SPA routing, package review using-grantpowers
A manuscript, not a proposal using-latexpowers

Floors every leaf enforces

  • Only verified claims are results. Published, submitted with the statement untouched by referees, or checked by the PI. Agent output the PI has not checked is "agents report …" or is held in REVIEW.md.
  • Prose for the PI's results; environments for the panel's anchors. Main definitions in definition, main conjectures in conjecture, at most two theorems; the PI's own results in prose with citations (PD-POLICY P2–P3).
  • Exact, sourced numbers. Every number has a REVIEW.md row with its source and date.
  • Register. Third person ("the PI", "this proposal"); direct, plain; no hype, no sentence whose only content is importance, no run-in labels or step lists (P1, P4, P5, P12). "AI-assisted, human-driven".
  • Define for the panel. Every term a non-specialist panelist would not know is defined once, where first used, in one sentence.
  • Show the data. Prefer a small table of real, verified records over a description of a schema.
  • No process in the mathematics. "We initially believed", earlier drafts, and model chatter belong only in the AI statement.

Finish

After each batch of edits: build; run purge-claudisms, check_pd_policy.py, check-grant-style and check-grant-format; confirm the page limits from the solicitation; update REVIEW.md; report what changed, what is unverified, and what the PI must decide. Never commit, push, or submit unless the PI asks.

04 同仓库更多 Skills